Skip to content
NewsModel release

DeepSeek V4.1 Flash ships as a 552B MoE with FP4 KV cache, MIT open weights, and Pro-tier routing on September 14

· by Pondero Newsdesk

The short version

DeepSeek released V4.1 Flash on September 10 with a 1M-token context window, an FP4 KV cache that drops per-token memory to 890 bytes, cross-layer attention reuse, native multimodal, and MIT open weights in 48 Hugging Face shards. V4 Pro API calls reroute to it at Flash prices starting September 14.

DeepSeek V4.1 Flash ships as a 552B MoE with FP4 KV cache, MIT open weights, and Pro-tier routing on September 14

A 552B mixture-of-experts model that fits the Flash pricing tier, carries a 1M-token context window, and compresses its KV cache to 890 bytes per token is the lead story in open-weight model releases this week. DeepSeek dropped V4.1 Flash on September 10, following an early-access window that opened September 9, and immediately put V4 Flash and V4 Flash Vision Exp into retirement under the new model name.

What shipped

V4.1 Flash uses a Causal Encoder-Decoder (CED) architecture. The Hugging Face README reports activation of 8B parameters per token during prefill and 16B during decode out of the 552B backbone, per ccleaks.com's coverage of the open-weight card.

The FP4 KV cache is the most concrete architectural advance. Quantized to E2M1 format under the NVFP4 standard, it brings the global KV cache footprint to 890 bytes per token. That is roughly one-quarter the footprint of V4 Flash and about 437 times smaller than DeepSeek-V1, per MarkTechPost's technical summary. Smaller KV cache means longer effective contexts and lower memory pressure at scale without changing the quality-compute tradeoff in the way that quantizing weights typically does.

Cross-layer attention reuse (Compressed Sparse Attention 2, or CSA2) runs in three modes: Full layers compute fresh attention, Reindex layers reuse KV while rescoring query representations, and Reuse layers skip indexing entirely. The result is lower attention compute at long-context lengths.

Multimodal capability is built into the base architecture, not bolted on through a vision adapter. Pre-training covered 45 trillion multimodal tokens at a 7:1 text-to-multimodal ratio, per MarkTechPost. This replaces the separate V4 Flash Vision Exp model, folding image understanding into the standard model string. GPQA Diamond sits at 90.9 in the benchmarks published with the release, per the DeepSeek changelog.

Open weights are on Hugging Face under the MIT license, ungated, in 48 safetensors shards. The API model name is deepseek-flash; legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route to V4.1 Flash, per ccleaks.com.

Why it matters

The September 14 routing change is the immediate decision point for API operators. At 12:00 Beijing Time on September 14, all requests to deepseek-v4-pro will route to V4.1 Flash and bill at Flash prices until V4.1 Pro ships, per DeepSeek's changelog. That means Pro-tier callers get a lower bill for the same model string without a configuration change. Any operator with cost controls tuned to Pro pricing should verify their budget assumptions before that date; the change lowers costs, but it also means the model behind deepseek-v4-pro will change silently.

For teams evaluating self-hosted deployments, the 48-shard MIT release answers the licensing question directly. The Causal Encoder-Decoder architecture requires FP4-capable hardware to realize the KV cache advantage. Teams running on H100s or B200s with NVFP4 support can exploit the 890-byte-per-token figure; teams on older cards will not see the same per-token memory gain and should benchmark on their own hardware before scaling.

Compared to third-party multimodal models with bolted-on vision adapters, the base-architecture integration in V4.1 Flash simplifies the deployment stack. One model serves both text and vision workloads instead of two separate model endpoints with different routing logic.

What to watch next

V4.1 Pro has no announced release date. Until it ships, V4 Pro API callers are on V4.1 Flash, which raises the practical question of whether V4.1 Flash meets the quality bar Pro-tier workloads require. Third-party benchmark comparisons of V4.1 Flash's FP4 KV cache efficiency against FlashAttention-3 on the same hardware will determine whether the memory savings hold up under production token-throughput conditions. The DeepSeek changelog is the first place updated numbers will appear.

Sources